Some researchers argue "hallucination" is the wrong word for what AI does. They prefer "confabulation" - a term from neuroscience for when the brain fills gaps in memory with fabricated details, without intending to deceive. The terminology debate is interesting. It changes nothing about the problem or the fix.
"Hallucination" is the established term. It is what the industry uses, what research papers use, what users search for, and what media reports on. It describes the observable behaviour accurately: the AI generates content that is not real and presents it as if it is.
The word works the same way it does for a human making up nonsense. A person who hallucinates sees something that is not there and believes it is real. An AI that hallucinates generates information that is not true and presents it as fact. The analogy holds.
In neuroscience, confabulation specifically describes gap-filling with fabricated memories. The patient is not lying - they genuinely believe their fabricated memory is real. They are filling a gap in knowledge with something that feels right.
This maps closely to what AI does. The model has a gap in its knowledge. It fills the gap with the most plausible continuation based on patterns. It is not "seeing things that aren't there" (hallucination). It is "filling memory gaps with fabricated content that it believes is correct" (confabulation).
Researchers who prefer this term argue it is more technically precise and avoids implying that AI has subjective experience or perception.
The AI never invents. It says what it thinks is right. Whether you call that hallucination or confabulation, the cause is the same: nothing forced it to verify.
Both terms describe the same observable behaviour: the model generates false content and presents it as true. Both have the same root cause: the absence of a verification mechanism.
Whether you call it hallucination or confabulation:
The model still predicts tokens probabilistically. It is still pattern-matching, not truth-checking.
The model still has no connection to reality. Its "truth" is still whatever patterns it learned from training data.
The fix is still verification. Feed the model live reality. Validate its output against that reality. The terminology changes. The engineering does not.
We use "hallucination" because it is the standard term. But whatever you call it, Check solves it the same way: real-time context injection and output validation against the live environment.
Some in the industry frame hallucination as a feature - the model's "creativity." The ability to generate novel content is what makes language models useful. Hallucination is just creativity that happens to be wrong.
This is marketing copywriting, not science. Calling a failure mode a feature is repositioning. It is "this is what it does now, so bear with it." It does not change the fact that when you ask an AI for the correct medication dosage and it gives you a fabricated one, that is not creativity. That is a dangerous failure.
The generative capability is genuinely useful. The inability to distinguish its own generation from verified reality is genuinely dangerous. These are not the same thing. One is the feature. The other is the missing grounding.
120 verifications a day free. No card, no signup.